基于改进型LeNet-5的工业机器人工件自动识别研究  

Research on automaticrecognition of industrial robot workpiece based on improved LeNet-5

作　　者：刘东来 崔亚飞 罗辉[1] 邓子林[1] 秦润华[1] 秦长江[2] LIU Donglai;CUI Yafei;LUO Hui;DENG Zilin;QIN Runhua;QIN Changjiang(Yongzhou Vocational Technical College,Yongzhou 425100,CHN;Xiangtan University,Xiangtan 411105,CHN)

机构地区：[1]永州职业技术学院,湖南永州425100 [2]湘潭大学,湖南湘潭411105

出　　处：《制造技术与机床》2021年第8期103-107,共5页Manufacturing Technology & Machine Tool

基　　金：湖南省教育厅科学研究项目(20C1868);湖南省科技厅科技计划重点研发项目(2016GK2014)。

摘　　要：针对机器人关节工件组装生产过程中,工件种类多、产量大、人工分拣与装配耗时费力等问题,在经典LeNet-5模型基础上,提出一种改进型LeNet-5网络,该网络输入图像的大小修改为32×32,卷积层增加至4层,激励函数改用Leaky ReLU以防止过拟合。同时,将改进型LeNet-5与经典LeNet-5、GoogLeNet模型进行训练、测试与对比,试验结果表明,改进型LeNet-5效果最好,测试集的准确率达到98.32%、曲线下降面积AUC为0.9163,识别一个待装配工件仅需约0.016 s,满足工厂工业机器人实时性检测要求,为类似的识别提供了有效参考,具有较高的应用价值。In the production process of robot joint workpiece assembly,there are some problems,such as many kinds of workpieces,large output,time-consuming and laborious as manual sorting and assembly,etc.Based on the classic LeNet-5 model,this paper presents an improved LeNet-5 network,which can modify the size of inputting image to 32×32,increase the convolution layer to 4 layers,and use Leaky ReLU instead of the excitation function to prevent over-fitting.Meanwhile,the improved LeNet-5 is trained,tested and compared with classic LeNet-5 and GoogLeNet models.The experimental results show that the improved LeNet-5 has the best effect,the accuracy of the test set reaches 98.32%,the loss value of AUC is 0.9163,and it only takes about 0.016 s to identify a workpiece to be assembled.It could meet the real-time detection requirements to industrial robots in factories,and provides an effective reference for similar identification,which has high application value.

关 键 词：工件识别 改进型LeNet-5 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]